CHANGELOG.md

v190.30 — Validation engine startup hardening

• Writes validation_engine.json and validation_preflight.json before fragile Gradio discovery/payload work.
• Guarantees validation_failure_diagnosis.json when linked validation stops before payload resolution.
• Fixes replay payload_source provenance to parent_automatic_smoke when backend replay source is present.
• Writes terminal summary.json for linked validation success/failure, avoiding stale running summaries.

v190.30 Eval archive run-scoped reset and layout stability

• v190.30 makes the Active Run eval archive card strictly run-scoped: launching, deleting, or switching Build Runs clears stale eval archive state before the selected run snapshot arrives.
• The eval archive key/value layout now uses a stable single-column grid with safe wrapping so long bucket names and record paths never collapse into letter-by-letter columns.

v190.30 — Runs Explorer linked validation parent metadata

• Build Run cards now show linked/manual validation strips from parent metadata (manual_validation_status and linked_validations) even when the child Validation Run is not present in the Runs Explorer cache.
• Runs Explorer remains build-first: linked validations stay summarized on their parent Build Run, while automatic Build Run smoke results remain separate from manual linked validation state.

v190.30 — Space Test payload provenance and prefill correctness

• Uses automatic smoke artefacts as the visible Space Test prefill source when available.
• Clears stale dirty validation args when a Build Run is selected from Runs Explorer.
• Sends and stores a backend validation launch payload for linked validations.
• Writes validation_launch_payload.json in validation runs before the worker resolves/predicts.
• Surfaces payload source provenance across UI, backend launch metadata, and worker artefacts.

v190.30 — Replay uses parent smoke artefact as source of truth

v190.30 — Linked validation replay hardening

• In Space Test replay mode, the validation worker now loads the parent Build Run's successful automatic smoke artefacts (generation_smoke.json, generation_smoke_payload*.json) as the canonical request before schema reconstruction.
• Added replay_source.json and richer resolved_validation_request.json / failure diagnosis fields so replay failures can be compared against the parent smoke request.
• The Payload stage now records Payload replayed from parent automatic smoke instead of remaining pending when replay data exists.

v190.30 — Linked validation worker schema coercion and replay fix

v190.30 — Validation worker reliability

• Normalizes linked validation replay payloads against the current Gradio endpoint schema before calling predict().
• Preserves typed Gradio choices and converts dropdown/radio numeric-looking args such as 1024 to schema strings such as "1024" when required.
• Writes validation_preflight.json, resolved_validation_request.json, schema_coercion.json, and validation_failure_diagnosis.json for every linked validation run.
• Keeps Space Test copy linked-only while explaining that the separate validation run remains attached to the selected Build Run.

v190.30 — UI/CSS convergence and Space Test validation polish

v190.30 — Space Test validation reliability

• Treats Space Test Validation as a first-class linked validation workflow.

• Adds a single resolved launch payload contract so replay/complete/recover/discovery-first modes only launch when coherent.

• Separates validation context from manual validation run state in the UI.

• Improves launch readiness messaging, payload-source visibility and advanced payload layout.

• Preserves endpoint discovery and schema-choice retry logic for linked validations.

• Removed transient legacy timeline rendering in favor of canonical placeholders.

• Cleaned Live test details to avoid stale endpoint-waiting rows and duplicate generation success text.

• Rebuilt Eval archive details as a readable key/value block.

• Polished linked-only Space Test validation context with clear endpoint, target Space, smoke latency, and launch labels.

v190.19 — Eval archive effective outcomes + owner flush

• Eval records now preserve automatic outcome and add effective outcome metadata for linked Space Test validations.
• Backend archive publishing enriches records with linked validation summaries, endpoint discovery metrics, smoke retry metrics and stricter privacy flags without storing prompts, args, target Space IDs or generated code.
• Eval publish statuses now use schema v1.3 and republish when a parent Build Run gains new effective validation state.
• Instance owners / configured eval admins can flush archived eval records from the mounted eval bucket while preserving archive configuration.

v190.19 — Build completion matrix + endpoint discovery-first Space Test

• Added canonical space_test_policy for replay/complete/recover/blocked/unavailable Build Run states.
• Linked Space Test can now discover Gradio endpoints before validation when a Build Run has no known endpoint schema.
• Space Test UI consumes policy labels/messages and shows discovery-first validation state.
• Runs Explorer/Stats can surface recoverable and needs-validation states without treating Space Test as standalone.

v190.19 — Linked validation hardening and effective status UX

• Fixed linked Space Test parent update after smoke success by sourcing api_name and expected_output_type from the smoke payload/environment instead of out-of-scope variables.
• Prevented failed linked validation attempts from downgrading a parent Build Run that already has a successful linked validation.
• Clarified effective status filters/stats, Eval Archive effective hints, Space Test context, and latency/ZeroGPU recommendation provenance.

v190.16 — Cascade delete linked Space Tests

• Deleting a Build Run now cascades to its linked Space Test validation runs so no hidden validation artifacts are orphaned.
• Deleting a linked validation run updates the parent Build Run linked validation metadata and manual validation status.
• The delete modal now reports linked validation counts and the frontend removes parent/child cache entries together.

v190.16 — Build-first explorer compatibility cleanup

• Keeps Runs Explorer simple after the linked-only Space Test change: Build Runs remain the primary list items, linked validations are summarized on their parent, and no standalone/legacy validation section is introduced.
• Hardens the run cache, stats, pagination, and selection paths against null, partial, or legacy Build Run records that do not contain manual validation metadata.
• Fixes linked validation parent selection without duplicating variables in the snapshot path.
• Keeps the v190.12 generic Gradio schema-choice smoke retry intact.

v190.16 hotfix — Runs Explorer null-safe linked validation cache

• Hardened Runs Explorer against null or malformed run summaries returned by stale bucket/list responses.
• Preserved build-first linked validation grouping while filtering invalid cache entries safely.
• Added regression coverage for null-safe linked validation parent lookup.

v190.16 — Build-first explorer and linked validation stats polish

• Treat linked validation runs as children of Build Runs in Runs Explorer by default.
• Rework Run Stats around Build Runs, linked tests, automatic success, and effective success.
• Stop the Active Run elapsed clock for partial/completed-with-warnings/manual-validated terminal states.
• Add softer Runs Explorer CSS spacing with borderless cards and linked-validation chips that breathe.

v190.13 — Linked-only Space Test and parent Build validation update

• Disabled standalone Space Test validation in the product UI/API path; validations must now be linked to a completed Build Run target Space.
• Prefill Space Test now carries a parent Build Run id and locks the target Space to the generated Space from that parent.
• Linked validation Jobs persist parent_build_run_id, source_kind, and linked target metadata.
• Successful linked validations write manual_validation_status.json and linked_validations.json back into the parent Build Run folder.
• Active Run snapshots and run summaries can now surface manual validation status, effective verdict, latency, and ZeroGPU duration measurements sourced from linked Space Test.
• Runs Explorer labels linked validation runs, shows manual validation passed on Build cards, and opens the parent Build Run when a linked validation is selected.

Agentic Space Factory v190.16

v190.16 smoke retry and final summary hardening

• Generation smoke payloads now preserve Gradio schema choice value types before calling predict.

• If Gradio rejects a smoke payload with a choice type mismatch, the worker records the first error, coerces args to schema choices, and retries once.

• Final worker completion now writes a terminal summary.json so lightweight run cards and bucket audits do not stay stuck on running.

• The worker writes live_status.json for micro-stage telemetry during live validation and smoke retry.

• Live Validation surfaces schema-choice retry details and next actions.

• Gated the Active Run “Prefill Space Test” action until a build reaches a final exploitable Space state (success, partial, or manual-action-required with a target Space).

• Added live validation telemetry to the canonical timeline model: Space runtime stage, health, Gradio schema/endpoints, generation smoke state, severity, and next action.

• Removed the redundant running subtitle “The pipeline is still in progress.” from the progress header.

• Tightened visible polling cadence for active build and validation runs so events feel less batched while preserving terminal stop/final refresh behavior.

Agentic Space Factory v190.10

• Live polling now uses the same canonical run snapshot as run selection/reload, so terminal builds converge across Active Run, Runs Explorer, Run Stats, and Eval Archive without a page refresh.
• Intermediate fallback/wait/test failures no longer mark the whole build as failed unless the canonical timeline is terminal.
• Initial build launch shows a quiet snapshot wait state instead of rendering the legacy full vertical timeline.
• Current step meta was removed; the canonical timeline owns phase context.
• Timeline phase details are hidden on normal happy-path phases and shown only for warnings, fallback, recovery, manual actions, archive issues, or terminal diagnostics.

Agentic Space Factory v190.9

• Persists/canonicalizes terminal Space Test validation status across page reloads so terminal artifacts/events override stale running launch metadata.
• Prevents HF Job stage or old summary data from resurrecting failed/success/partial validations as running in /api/runs and /api/runs/{id}/progress.
• Makes the delete confirmation modal use the same canonical validation status, so failed validations no longer warn that they still look active.
• Keeps validation auto-resume/polling from restarting terminal validations after reload; explicit “Run validation again” remains the only relaunch path.

Agentic Space Factory v190.6

• Locks terminal Space Test validation states so stale running/pending data cannot override failed/success/partial/manual outcomes.
• Keeps validation panels stable during slow refreshes using stale-while-revalidate semantics.
• Preserves terminal validation status in Runs Explorer while backend summaries refresh.

Agentic Space Factory v190.5

Active Run timeline + Space Test prefill hotfix

• Restores the canonical Active Run timeline as a horizontal fit-to-width row across narrow central columns.
• Keeps the active timeline phase pulsing while a phase is running.
• Ensures the Agent phase remains running while pi_run is started or running, instead of prematurely showing complete.
• Improves Space Test prefill from partial/failed builds by reading nested inference_gate.generation_smoke payloads and endpoint parameters when available.
• Keeps the validation prefill in new-validation draft mode rather than reusing a previous validation selection.

v190.5 — Space Test explicit prefill mode

• Fixed Active Run → Space Test prefill so it prepares a new validation draft instead of re-opening the previous validation run.
• Clearing a prefilled Space Test now clears selectedValidationId/validation polling state and disables delete until a new validation is launched.
• Existing validation cards in Runs Explorer still open in view-existing-validation mode.
• Added regression tests for explicit validation prefill vs existing validation selection.

v190.5 — Active Run timeline hotfix

• Restores the Active Run canonical timeline as a horizontal row on desktop/tablet after the fit-to-width consolidation.
• Lowers the horizontal breakpoint to avoid accidental vertical stacking in narrow center columns while keeping a true mobile fallback.
• Restores the running-step pulse/ring animation on the active canonical timeline node.
• Keeps pi_run started/running as Agent running instead of marking Agent complete because earlier setup events succeeded.
• Leaves Space Test and worker/runtime behavior unchanged.

v190.1 — Space Test Validation UI/State Model

• Adds a dedicated canonical Space Test timeline: Start, Runtime, API, Endpoint, Payload, Smoke test, Result, Report. Validation views no longer reuse build/recovery phases such as repair upload, manual hardware, or upload files.
• Fixes Run Explorer status normalization for validation runs so terminal validation verdicts such as full_inference_success override stale running summaries.
• Humanizes Gradio endpoint schema rendering: component metadata objects are normalized into labels, choices/defaults are summarized, and [object Object] is never displayed.
• Adapts Space Test labels for completed validation runs: terminal runs can show Run validation again, latency displays Not recorded when missing, and running validations keep actions disabled.
• Keeps build worker/runtime behavior unchanged.

v190.0 — Active Run layout consolidation

• Consolidates the Active Run UI without runtime changes: canonical timelines now fit the center column on desktop, hide the redundant Done dot, and relabel Hardware as GPU.
• Keeps Run notes as the only compact home for non-blocking signals, with explicit Pi assistant mismatch / GPU fallback wording.
• Makes Recovery phase details compact when Agent recovery provides the full diagnosis, avoiding duplicate Diagnose/Decide/Patch/Revalidate rows.
• Improves Agent recovery text wrapping so blocker explanations and next actions are not truncated.

v189.9 — Timeline polish and Run notes clarity

UI polish

• Keeps Recovery phase details compact when the dedicated Agent recovery panel is available.
• Uses Run notes as the single compact place for non-blocking warnings such as Pi model routing and hardware fallback.
• Removes opaque +N more style note summaries from the canonical warning banner.
• Keeps timeline phase details as navigation summaries rather than duplicating specialized diagnostic panels.

Runtime

• No worker/runtime behavior changed from v189.8. The v189.7 guardrails remain: quoted safe python_version, torch added when imported, and .cache/ excluded from generated Space uploads.

v189.8 — Runtime Sanitation + Failed Timeline Truth

• Sanitizes README python_version to quoted safe values (3.10, 3.11, 3.12) so 3.10 is never interpreted as YAML float 3.1.
• Adds torch>=2.0.0 automatically when generated app.py imports torch but requirements omit it.
• Excludes transient .cache/ files from generated Space uploads.
• Makes canonical timeline archive state prefer confirmed published=true over stale record_not_ready events.
• Improves failed recovery summaries so final build blockers are surfaced distinctly from initial repaired blockers.
• Keeps timeline node labels compact in both canonical and legacy render paths.
• Prioritizes failed-run documents: Report, Blockage, Build error, Space logs, Decision, Repair.

v189.6 — Timeline State Hotfix

• Derives canonical phase status from the latest status per worker step, not any historical event.
• Prevents terminal runs from leaving early phases visually running because of older started events.
• Prevents resolved Space log collection warnings from over-marking Deploy when final log quality is full.
• Refines terminal active-phase selection so full-success runs focus Live test/Archive/Done while non-blocking warnings stay grouped separately.
• Keeps the validated worker/runtime path unchanged.

v189.4 — Mobile Timeline Polish

• Adds an Active Run phase details panel for the canonical run_timeline_model.v1.
• Uses a mobile-first vertical canonical timeline on narrow screens.
• Shows compact phase details and Needs attention summaries without changing the validated runtime worker path.
• Keeps legacy timeline fallback behavior for older runs without timeline_model.

v189.2 — Canonical Timeline UI bridge

• Active Run now uses run_timeline_model.v1 for the progress header and main timeline whenever the backend view model provides it.
• The legacy compact worker timeline remains as a fallback for old or partial payloads.
• Timeline phases now represent product-level phases instead of re-inferred worker step groups, reducing conflicting progress states.

v184 — Compact Model Pre-scan Card

• Refined the default Model pre-scan card to be more visible without adding heavy borders.
• Reworked successful model pre-scan results into a compact premium summary with metadata icon chips, inline signal pills, and a single recommendation line.
• Reduced the visual footprint of safe scans while preserving pipeline, library, expected output, good signals, and recommendations.

v176 — startup import fix for eval archive build

• Fixed a syntax error in src/bucket.py caused by nested quotes inside the launch summary artifacts_url f-string.
• Verified that the app imports/compiles cleanly again and that the full test suite passes.

v174 — Bucket run-prefix mount and robust worker bootstrap writes

• Mounted the runs/ Storage Bucket prefix explicitly into Jobs using Volume(type="bucket", source=..., path="runs", mount_path="/output/runs").
• Preserved the worker contract: OUTPUT_ROOT=/output, run_dir=/output/runs/<run_id>, and remote objects under runs/<run_id>/.
• Hardened API-side _worker.py persistence with HfFileSystem.write_text, pipe_file, open(..., "wb"), and a final sync_bucket fallback.
• Removed virtual-directory creation for Bucket prefixes and kept deletion/listing/document links aligned with the configurable run prefix.
• Raised the minimum huggingface_hub requirement to >=1.11.0.

v173 — Bucket path separation fix

• Reverts the v172 API-side makedirs workaround for fresh run prefixes.
• Keeps a strict distinction between remote bucket object URIs (hf://buckets/<namespace>/<bucket>/runs/<run_id>/...) and Job mount paths (/output/runs/<run_id>/...).
• Writes API-side bucket objects directly with HfFileSystem.pipe_file(...), with open(..., "w") as a compatibility fallback.
• Adds tests to prevent treating virtual runs/ prefixes as directories/repositories.

v172 — fresh bucket writes and pre-scan polish

• Keeps Hugging Face Space metadata at the very top of README.md so the app archive cannot boot with a configuration error.
• Makes fresh Bucket run-prefix writes more robust by creating parent prefixes before writing _worker.py, launch.json, summary.json, state.json, and other API-authored artifacts.
• Restyles the Model pre-scan card into a calmer premium status panel with compact metadata chips and less oversized success treatment.

v171 — delete Space option, ghost-run cleanup, storage polish

• Added an opt-in checkbox in the premium delete modal to delete the associated generated Space for build runs.
• Kept validation runs protected from accidental associated Space deletion.
• Made ghost/unknown run prefixes deletable from Run Explorer.
• Compacted the Run Storage setup error card.
• Preserved the v170 real recursive bucket deletion behavior.

v120 — Agent recovery UI

• Added the Agent recovery section to summarize Pi recovery decisions, blockers, evidence, and repair/revalidation status.
• Kept detailed source files in Run traces instead of duplicating raw recovery logs in the summary card.